University of Massachuses Amherst
From the SelectedWorks of Jianhua Yang
$$'-%)
&+'%/'&'*$'&,+!'& &&$
-$*+ &!*%'! )')%&
%)!*+')
&!'
' &+) &
!& ,& University of Massachuses - Amherst
&
$&'$)+$
-!$$+ 0(*.')#*()**'%"!& ,/&
1
Present memory technologies, including DRAM (dynamic
random access memory), SRAM (static random access memory),
and fl ash, are potentially approaching their scalability limits
near the 16 nm technology generation.
[
1–3
]
Accordingly, the Inter-
national Technology Roadmap for Semiconductors (ITRS) has
recently completed an assessment of eight memory technologies
among the emerging research devices (ERDs) and recommended
that Redox RAM
[
4–7
]
and STT-MRAM
[
8
,
9
]
(spin-transfer torque
magnetic RAM) receive additional focus in research and devel-
opment.
[
1
]
Redox RAM is one type of memristor
[
10–15
]
that has
shown more than adequate scalability, non-volatility, multiple-
state operation, 3D stackability, and complementary metal-
oxide semiconductor (CMOS) compatibility.
[
16–39
]
Moreover,
these devices have also exhibited signifi cant potential in other
applications, such as stateful logic operations,
[
40
]
neuromorphic
computing,
[
27
,
41
]
and CMOS/memristor hybrid circuits for con-
fi guration bits and signal routing.
[
42
]
However, as pointed out
by the ITRS 2010, there are still challenges remaining for these
devices, among which are reliability and a better understanding
of the microscopic picture of the switching.
[
1
]
The reliability
issue includes switching endurance, long-term thermal sta-
bility, repeatability, and robustness of the conduction channel,
[
1
]
all of which depend on the chemical and structural details of
the conduction channel revealed in this study.
Signifi cant progress has been made in device performance
using tantalum oxide based memristors. We recently demon-
strated over 10
10
open-loop switching cycles
[
43
]
in tantalum oxide
based devices, with the device remaining switchable after 15 bil-
lion cycles without any feedback or power-limiting circuits.
[
43
,
44
]
This endurance record was surpassed within a year to 10
12
by Lee et al.,
[
45
]
also using a tantalum oxide system. In addi-
tion, switching in less than 2 ns is demonstrated in Figure 1 a,
using a relatively low voltage ( < 2 V) for switching both ON
(set) and OFF (reset). The as-prepared device was activated into
normal switching operation by a 5 ns pulse with the same ON
switching voltage utilized in the subsequent switching cycles
(inset to Figure 1 a), and so no high voltage or long-time elec-
troforming was required. Favorable device size scaling is dem-
onstrated in Figure 1 b,c, in which crossbar nanodevices with
50 nm half-pitch (device size and also the distance between two
neighboring devices) show less than 10 μ A switching current,
which is encouraging for low-energy device applications. These
promising device characteristics highlight the importance of
understanding both the structure and switching mechanism
of tantalum oxide based devices in order to engineer further
improvements.
Switching mechanisms in many transition metal oxide based
resistance switches
[
5
,
6
,
13
,
46
]
have been extensively studied, and
the drift/diffusion of oxygen vacancies
[
47
]
(or anions) driven by
an electric fi eld and/or thermochemical reduction/oxidation
[
48
]
are believed to play a key role. Lee et al.
[
45
,
49
]
have adopted a
drift switching mechanism similar to that introduced for
TiO
2
memristive switching
[
13
]
to explain tantalum oxide based
devices. However, there are strong qualitative differences
[
43
]
in
the switching observed between Ti-O and Ta-O systems, sug-
gesting major differences in the mechanism that should be
observable. Thus, it is crucial to identify and image the actual
“active” switching region buried inside the device since elec-
trical operation, especially any electroforming process (used
in ref. [ 45 ] ), can signifi cantly change the as-fabricated struc-
ture and material composition. This has been a longstanding
desire since the 1960s
[
4
,
50
,
51
]
and important progress has been
made recently.
[
46
,
52
,
53
]
For example, a crystalline sub-oxide has
been identifi ed as the conduction channel in both unipolar
and bipolar TiO
2
-based switches.
[
46
,
54
]
One approach is to pre-
pare and examine a large number of sample slices through
a device that has been switched in order to fi nd one with
fi lament-like features. This technique is time-consuming and
cannot guarantee success in fi nding “the” active conduction
channel responsible for the switching behavior observed in a
device before it is destroyed. Here, we implemented a method for
locating the nanoscale channel responsible for bipolar switching
in a Pt/TaO
x
/Ta memristor. A microscopic picture of the con-
duction channel was then obtained, which was different from
that observed for Ti-O and signifi cantly constrained the possible
switching mechanism of the tantalum oxide system. We propose
a new switching mechanism for the tantalum oxide memristor
system that is consistent with both the structure of the conduc-
tion channel and the observed electrical properties of the device.
As shown in Figure 2 , the method we employed for
fi nding and characterizing the active conduction channels
Feng Miao , John Paul Strachan , J. Joshua Yang , Min-Xian Zhang , Ilan Goldfarb , Antonio
C. Torrezan , Peter Eschbach , Ronald D. Kelley , Gilberto Medeiros-Ribeiro , and
R. Stanley
Williams
Anatomy of a Nanoscale Conduction Channel Reveals
the Mechanism of a High-Performance Memristor
2
device at a small current bias, yielding a resistance map as a
function of tip position. When the tip pressurized an active
region in which the fl ow of current was dominant, a resistance
dip appeared on resistance maps, originating from the com-
pression of the conduction channel
[
56
]
or decrease of the tun-
neling or hopping distance.
[
55
,
57
]
By correlating the resistance
map with the AFM topography image collected simultaneously,
we were able to locate the position of the buried active switching
region with a lateral resolution determined by the radius of the
AFM tip ( ≈ 10 nm). Figure 2 b shows the PMCM resistance map
of a tantalum-oxide-based memristor (see Experimental Section
consisted of three steps: i) precisely locating the active switching
region(s) through pressure-modulated conductance microscopy
(PMCM)
[
55
]
on a functioning device; ii) cross-sectioning the
active switching region(s) by focused ion beam (FIB) milling;
and iii) examining the structure and composition through high-
resolution cross-sectional transmission electron microscopy
(X-TEM) and electron-energy-loss spectroscopy (EELS).
The PMCM procedure is schematically shown in Figure 2 a.
We used a non-conducting atomic force microscopy (AFM)
tip to apply pressure to the top electrode of a cross-bar device
and simultaneously monitored the change of resistance of the
Figure 1 . Electrical performance of tantalum oxide based memristors. a) Repeatable high speed switching under 2 ns measured in real-time for ON
and OFF switching with a less than 2 V pulse (showing ten voltage–time curves for each). No signifi cant electroforming is required: lower left inset
shows the fi rst ON switching, which activated the as-prepared device into a switchable state using a 5 ns, < 2 V pulse. Upper right inset shows the
reproducibility of the resistance values for the fi rst ten ON and OFF pulses. b) 17 × 17 memristor cross bar with 50 nm half pitch. c) Current–voltage
curves from a 100 μ m diameter disc device and a 50 nm half-pitch crosspoint device (inset), showing the fabrication scalability and decreasing current
level with decreasing device size.
3
dark-fi eld imaging around the location of the channel, a 5 nm
nanocrystal was observed in Figure 3 a. As grown, the tantalum
oxide layer was amorphous and high-resolution TEM (HRTEM,
Figure 3 b) showed that the switching layer in the channel region
remained amorphous throughout, except for this nanocrystal.
To identify its composition and structure, convergent beam
electron diffraction (CBED, bottom-left inset in Figure 3 b) was
performed and revealed only two intense diffraction spots cor-
responding to an interplanar spacing of 1.81 Å. We identifi ed
this structure as a base-centered monoclinic variant of a high-
temperature (HT) tetragonal
α
-Ta
2
O
5
phase.
[
58
]
While the above
measured interplanar distance could also match that of
β
-Ta
2
O
5
phase, the particular symmetry and extinction conditions, with
only one set of strongly refl ecting planes in the relevant zone,
were more consistent with the HT
α
-Ta
2
O
5
phase, as was cor-
roborated by the diffraction simulation. The corresponding
simulated (001) SAED (selected area electron diffraction), with
diffracting 1.90 Å-spaced {020} planes, is shown in the right-
bottom inset in Figure 3 b. This phase is obtained by distortion
of the HT
α
-Ta
2
O
5
phase,
[
58
]
and the transformation from a low-
temperature (LT) orthorhombic
β
-Ta
2
O
5
to the HT tetragonal
α
-Ta
2
O
5
phase
[
58
]
requires a temperature of 1633 K in bulk sam-
ples.
[
59
,
60
]
While much lower temperatures were likely required
for crystallization from the as-grown, nanoscale amorphous
TaO
x
fi lm, this still suggests that a relatively high tempera-
ture was attained in this localized region. Thus the presence
of the insulating nanocrystal provides evidence of signifi cant
heating during the electrical operation, most likely caused by
Joule heating from a nearby conduction channel, which makes
temperature an important component of the switching mecha-
nism. In addition, the known stoichiometric composition of the
crystallite provides a good internal calibration for determining
the Ta to O ratio around the channel region.
EELS was used to study the elemental composition around
the channel region of Figure 3 c. Horizontal and vertical line
scans were collected, as indicated by the dashed lines, with the
corresponding oxygen profi les shown in Figure 3 d,e. By com-
paring this to the oxygen signal in a region of the device far
away from the channel region (dashed black line of Figure 3 e
labeled “pristine”), it is clear that the oxygen content signifi -
cantly decreased with lateral position from the Ta
2
O
5
nano-
crystal toward the center of the channel region, and the oxygen
content inside the channel region was less than one third of
that in the pristine fi lm or the nanocrystal. Thus, the conduc-
tion channel consisted of Ta(O) solid solution and was amor-
phous (Figure 3 b). This is in contrast to the crystalline con-
duction channels (Magneli Ti
4
O
7
) found in the titanium oxide
system,
[
46
,
54
]
which has a signifi cantly more complex phase dia-
gram. As shown in Figure 3 d,e, there was a distinct horizontal
oxygen gradient and a much less pronounced vertical oxygen
gradient within the channel region. The oxygen content was
surprisingly high in the boundary region between the crystal
and the conduction channel, as shown by the yellow dashed
scan line in Figure 3 c,e in addition to Figure 3 d. In contrast,
the Ta signal (not shown) did not show a signifi cant variation.
Important complementary information on the nature of the
conduction channel was recently provided by hard X-ray spec-
tromicroscopy
[
61
]
(Figure 3 f) performed on a Pt/TaO
x
/Pt device,
which exhibited similar switching behavior as the Pt/TaO
x
/
for details) in a low resistance state (ON state). On the resist-
ance map, only one resistance dip of approximately 100 nm
diameter (yellow–red dot) was observed, revealing that only one
active switching region existed or was dominant in this device.
Once the active switching region was located by PMCM, a dual
beam FIB/scanning electron microscope was used to cross-
section it, using careful registration to the PMCM map in
order to cut across the identifi ed channel. A TEM image of
the resulting cross-section that includes the channel region is
shown in Figure 2 c. The structure and elemental composition
of this area were then analyzed in detail.
Figure 3 shows the complete physical characterization of the
conduction channel region from electron microscopy, spectros-
copy, and diffraction (Figure 3 a–e), and focused X-ray analysis
from a synchrotron on a separate device (Figure 3 f). Performing
Figure 2 . Identifi cation and visualization of the conduction channel.
a) Schematic illustration of PMCM, for which a non-conducting AFM tip
applied pressure to the top electrode while the resistance of the device
was monitored, yielding a resistance map as a function of tip position.
b) The resistance map of a TaO
x
-based memristor, where the red dot
(resistance decrease), highlighted by the dashed square in the magnifi ed
inset, corresponds to the conduction channel. The color scale represents
the measured resistance values. The conduction channel was cross-
sectioned by FIB across the center (indicated by the black dashed line
in the inset). c) TEM image of the conduction channel region identifi ed
from PMCM.
4
comparable to that observed by PMCM (Figure 2 b). The higher
energy used in the very fi rst switching cycle due to the longer
pulse duration (Figure 1 a) combined with the initially different
heat dissipation condition are responsible for the formation of
the crystalline phases. The switching region in the subsequent
switching cycles is then likely contained within the surrounding
crystalline region because of the relatively low tolerance of off-
stoichiometry of crystalline materials.
The compositional changes and transport properties of
the conduction channel (Figure 3 ) were further studied by
growing and analyzing similar amorphous tantalum oxide
fi lms with various oxygen concentrations. The resistivity and
temperature coeffi cient of resistance (TCR) of the fi lms were
Ta devices. A Ta-rich region was identifi ed in the center of the
image in Figure 3 f, which corresponded to the Ta(O) solid solu-
tion phase described above. Surrounding this Ta(O) phase was
a differentiated Ta-oxide phase that, from analysis of the L-edge
absorption spectrum, had an increased short-range order
compared to the amorphous Ta
2
O
5
fi lm (outer gray region of
image), but without any increased concentration of Ta, therefore
matching the description of the crystalline HT
α
-Ta
2
O
5
phase
from TEM. While only one nanocrystal was found by TEM, the
larger area and volume probed by X-ray microscopy suggests
that more highly ordered and stoichiometric Ta
2
O
5
in fact sur-
rounded the channel. The dimension of the channel region
(around 150 nm) revealed by hard X-ray spectromicroscopy was
Figure 3 . Structural and compositional analysis by electron microscopy, diffraction, and spectroscopy of the channel region identifi ed in Figure 2.
a) Dark-fi eld imaging of the switching region showed a crystalline grain adjacent to the channel. SAED in the area provided several diffraction peaks that
were selected for this dark-fi eld image. b) HRTEM revealed a lattice pattern from a ≈ 5 nm nanocrystal, with the surrounding oxide being amorphous.
CBED (bottom left) was taken within the nanocrystal and indexed by simulation (bottom right) to be
α
-Ta
2
O
5
. c) Bright-fi eld TEM of the channel region
with dashed lines indicating where horizontal and vertical EELS line profi les were obtained. d,e) Horizontal and vertical line profi les, respectively, of
the background subtracted, integrated oxygen K-edge jump height from EELS, which is proportional to the oxygen atomic concentration. Horizontal
profi le (d) was collected across the border of the crystal and the channel region, showing a high oxygen concentration in the border region. Several
vertical profi les (e) were collected moving toward the central channel region away from the nanocrystal (white outline). The oxygen profi le for the pris-
tine tantalum oxide fi lm is shown (dashed black curve) for comparison. f) Synchrotron based nanobeam X-ray fl uorescence measurement of a similar
tantalum oxide device showing a central Ta-rich channel, surrounded by a nanocrystalline Ta
2
O
5
phase.
